Upload
others
View
3
Download
0
Embed Size (px)
Citation preview
Robust and Efficient Fitting of the Generalized Pareto
Distribution with Actuarial Applications in View
Vytaras Brazauskas1
University of Wisconsin-Milwaukee
Andreas Kleefeld2
University of Wisconsin-Milwaukee
Abstract
Due to advances in extreme value theory, the generalized Pareto distribution (GPD) emerged as a
natural family for modeling exceedances over a high threshold. Its importance in applications (e.g.,
insurance, finance, economics, engineering and numerous other fields) can hardly be overstated
and is widely documented. However, despite the sound theoretical basis and wide applicability,
fitting of this distribution in practice is not a trivial exercise. Traditional methods such as max-
imum likelihood and method-of-moments are undefined in some regions of the parameter space.
Alternative approaches exist but they lack either robustness (e.g., probability-weighted moments)
or efficiency (e.g., method-of-medians), or present significant numerical problems (e.g., minimum-
divergence procedures). In this article, we propose a computationally tractable method for fitting
the GPD, which is applicable for all parameter values and offers competitive trade-offs between
robustness and efficiency. The method is based on ‘trimmed moments’. Large-sample properties of
the new estimators are provided, and their small-sample behavior under several scenarios of data
contamination is investigated through simulations. We also study the effect of our methodology
on actuarial applications. In particular, using the new approach, we fit the GPD to the Danish
insurance data and apply the fitted model to a few risk measurement and ratemaking exercises.
JEL Codes : C13, C14, C16, C46. Subject Categories : IM10, IM11, IM41, IM54.Insurance Branch Categories: IB30.
Keywords : Pure Premium; Robust Statistics; Simulations; Trimmed Moments; Value-at-Risk.
1Corresponding author: Vytaras Brazauskas is Associate Professor, Department of Mathematical Sciences, Uni-
versity of Wisconsin-Milwaukee, P.O. Box 413, Milwaukee, WI 53201, U.S.A. E-mail address: [email protected] Kleefeld is Ph.D. Candidate, Department of Mathematical Sciences, University of Wisconsin-Milwaukee,
P.O. Box 413, Milwaukee, WI 53201, U.S.A. E-mail address: [email protected]
1 Introduction
Due to advances in extreme value theory, the generalized Pareto distribution (GPD) emerged as a
natural family for modeling exceedances over a high threshold. Its importance in applications such as
insurance, finance, economics, engineering and numerous other fields, can hardly be overstated and is
widely documented. For a general introduction to this distribution, the reader can be referred to, for
example, Johnson et al . (1994, Chapter 20) and McNeil et al . (2005, Chapter 7).
There is also a substantial literature on various specialized topics involving the GPD. For example,
the problem of parameter estimation has been addressed by Hosking and Wallis (1987), Castillo and
Hadi (1997), Peng and Welsh (2001), and Juárez and Schucany (2004), among others. The work of
Davison and Smith (1990) presents extensions of the GPD to data with covariates and time series
models. It also contains insightful discussions about the validity of the GPD assumption in real-world
applications. An interesting approach toward threshold selection has been proposed by Dupuis (1999);
it is based on robust procedures. Finally, the papers by McNeil (1997) and Cebrian et al . (2003) are
excellent illustrations of the GPD’s role in actuarial applications.
However, despite the sound theoretical basis and wide applicability, fitting of this distribution in
practice is not a trivial exercise. Traditional methods such as maximum likelihood and method-of-
moments are undefined in some regions of the parameter space. Alternative approaches exist and
include estimators that are based on: probability-weighted moments (Hosking and Wallis, 1987),
method-of-medians (Peng and Welsh, 2001), quantiles (Castillo and Hadi, 1997), and minimum-
divergence procedures (Juárez and Schucany, 2004). While each methodology has its own advantages,
it should also be pointed out some of their caveats. For example, the probability-weighted estimators
lack robustness (see the influence curve plots of Davison and Smith, 1990), the method-of-medians
estimators are not very efficient (see He and Fung, 1999, who introduced this method of parameter
estimation), and the minimum-divergence estimators can present significant computational problems
(see the simulation studies of Juárez and Schucany, 2004).
In this article, we propose a computationally tractable method for fitting the GPD, which is
applicable for all parameter values and, in addition, offers competitive trade-offs between robustness
1
and efficiency. The method is based on ‘trimmed moments’, abbreviated as MTM, and the resulting
estimators are found by following a general methodology introduced by Brazauskas et al . (2009). Note
that the idea of coupling robust-statistics methods and extreme-value data is not a contradiction. This
has been argued by some of the aforementioned authors, but perhaps the most eloquent discussion
on this topic has been provided by Dell’Aquila and Embrechts (2006). Interestingly, the estimators of
Castillo and Hadi (1997) also possess similar qualities as MTMs, though they have not been presented
or viewed from the robustness and efficiency perspective. As we will show later, the quantile-based
estimators are a special/limiting case of MTMs, which occurs when one trims all the available data
except for a few observations, i.e., the chosen quantiles. It is also worthwhile mentioning that in the
actuarial literature, the quantile-based estimators are better known as estimators that are derived by
employing the percentile-matching approach (see, Klugman et al ., 2004, Section 12.1).
The rest of the article is organized as follows. In Section 2, we first provide key distributional
properties of the GPD and graphically examine shape changes of its density function. Later, we
study various issues related to model-fitting. Specifically, a number of methods for the estimation of
parameters (e.g., method of maximum likelihood, method of moments, percentile-matching method,
and method of trimmed moments) are presented, and large- and small-sample robustness properties
of these estimators are investigated in detail. In Section 3, we fit the GPD to the Danish insurance
data. The fitted models are then employed in a few ratemaking and quantitative risk management
examples. In particular, point estimates for several value-at-risk measures and net premiums are
calculated. Results are summarized and conclusions are drawn in Section 4.
2 Fitting the GPD
In this section, we study essential issues related to model-fitting. The key facts and formulas of the
GPD are presented, illustrated and discussed in subsection 2.1. A number of existing and new methods
for estimation of the GPD parameters are provided in subsection 2.2. Finally, subsection 2.3 is devoted
to small-sample properties of the (theoretically and computationally) most favorable estimators, which
are investigated using simulations.
2
2.1 The Model
The cumulative distribution function (cdf) of the GPD is given by
F (x) =
{1 − (1 − γ(x − x0)/σ)1/γ , γ 6= 01 − exp (−(x − x0)/σ) , γ = 0,
(2.1)
and the probability density function (pdf) by
f(x) =
{σ−1 (1 − γ(x − x0)/σ)1/γ−1 , γ 6= 0σ−1 exp (−(x − x0)/σ) , γ = 0,
(2.2)
where the pdf is positive for x ≥ x0, when γ ≤ 0, or for x0 ≤ x ≤ σ/γ, when γ > 0. The parameters
−∞ < x0 < ∞, σ > 0, and −∞ < γ < ∞ control the location, scale, and shape of the distribution,
respectively. In insurance applications, the location parameter x0 is typically known and can be
interpreted as a deductible, retention level, or attachment point. But σ and γ are both unknown
and have to be estimated from the data. Note that, when γ = 0 and γ = 1, the GPD reduces to
the exponential distribution (with location x0 and scale σ) and the uniform distribution on [x0, σ],
respectively. If γ < 0, then the Pareto distributions are obtained. In Figure 1, we illustrate shape
changes of f(x) for different choices of the shape parameter γ.
0 0.5 1 1.5 20
1
2
3
x
f (x
) γ = 1.5
γ = 1.0
γ = 0.75
γ = 0.5
0 1 2 3 4 50
0.25
0.5
0.75
1
x
f (x
)
γ = −2 γ = −1 γ = 0 γ = 0.25
Figure 1: Probability density functions of the GPD for x0 = 0, σ = 1, and various values of γ.
Further, besides functional simplicity of its cdf and pdf, another attractive feature of the GPD is
that its quantile function (qf) has an explicit formula. This is especially useful for estimation of the
3
parameters via MTM and percentile-matching methods, and for portfolio risk evaluations by using
value-at-risk measures. Specifically, the qf is given by
F−1(u) =
{x0 + (σ/γ) (1 − (1 − u)γ) , γ 6= 0x0 − σ log(1 − u), γ = 0.
(2.3)
Also, the mean and variance of the GPD are given by
E[X] = x0 +σ
1 + γ, γ > −1 (2.4)
Var[X] =σ2
(2γ + 1) (γ + 1)2, γ > −1/2. (2.5)
Finally, as discussed by Hosking and Wallis (1987), GPDs with γ > 1/2 have finite endpoints with
f(x) > 0 at each endpoint (see Figure 1), and such shapes rarely occur in statistical applications.
GPDs with γ ≤ −1/2 have infinite variance (see expression (2.5)), and this too is unusual in statistical
applications. We certainly agree with this assessment of the statistical applications but in actuarial
work things are slightly different. For example, when pricing an insurance layer, it is not unreasonable
to employ a probability distribution with both endpoints finite. What is even more common in actu-
arial applications is the heavy-tailed distributions, with no finite moments at all, that often appear in
modeling of catastrophic claims. This discussion, therefore, suggests that parameter-estimation meth-
ods that work over the entire parameter space of the GPD are indeed needed in actuarial applications.
2.2 Parameter Estimation
In subsection 2.2.1, standard estimators, based on the maximum likelihood and method-of-moments
approaches, are presented and their large-sample properties are examined. Then, in subsection 2.2.2,
we briefly review quantile-type (percentile-matching) estimators and specify their large-sample distri-
bution. Finally, in subsection 2.2.3, we consider a recently introduced robust estimation technique,
the method of trimmed moments, and study asymptotic behavior of estimators based on it.
Also, throughout this section, we will consider a sample of n independent and identically distributed
random variables, X1, . . . ,Xn, from a GPD family with its cdf, pdf, and qf given by (2.1), (2.2) and
(2.3), respectively, and denote X1:n ≤ · · · ≤ Xn:n the order statistics of X1, . . . ,Xn.
4
2.2.1 Standard Methods
A maximum likelihood estimator of (σ, γ) of the GPD, denoted by (σ̂MLE, γ̂MLE), is found by numerically
maximizing the log-likelihood function:
logL(σ, γ |X1, . . . ,Xn) = −n log σ +1 − γ
γ
n∑
i=1
log(1 − γ
σ(Xi − x0)
).
Note that this function can be made arbitrarily large by choosing γ > 1 and σ/γ close to the maximum
of Xi’s, i.e., close to Xn:n. Hence, the obtained result will be a local maximum, not global. Further,
as presented by Hosking and Wallis (1987), the estimator (σ̂MLE, γ̂MLE) is consistent, asymptotically
normal, and asymptotically efficient when γ < 1/2. Specifically, we shall write that, provided γ < 1/2,
(σ̂MLE, γ̂MLE
)∼ AN
((σ, γ), n−1Σ0
)with Σ0 = (1 − γ)
[2σ2 σσ (1 − γ)
], (2.6)
where AN stands for ‘asymptotically normal’. On the other hand, if γ ≥ 1/2, then we have the
nonregular situation. The latter case has been treated by Smith (1985) who demonstrated that: for
γ = 1/2, MLE’s asymptotic distribution is still normal but with a different convergence rate; for
1/2 < γ ≤ 1, it is not normal (and very messy); and for γ > 1, the MLE does not exist. We will not
consider the nonregular case in this paper.
Next, a method-of-moments estimator of (σ, γ) can be found by matching the GPD mean and
variance, given by (2.4) and (2.5), with the sample mean X and variance S2, and then solving the
system of equations with respect to σ and γ. This leads to
σ̂MM = 0.5(X − x0
) ((X − x0
)2 /S2 + 1
)and γ̂MM = 0.5
((X − x0
)2 /S2 − 1
).
If γ > −1/4, then(σ̂MM, γ̂MM
)∼ AN
((σ, γ), n−1Σ1
)with
Σ1 =(1 + γ)2
(1 + 3γ)(1 + 4γ)
[2σ2(1 + 6γ + 12γ2)/(1 + 2γ) σ(1 + 4γ + 12γ2)
σ(1 + 4γ + 12γ2) (1 + 2γ)(1 + γ + 6γ2)
]. (2.7)
In the special case γ = 0, the MLE and MM are the same, which implies that the asymptotic
normality results (2.6) and (2.7) are identical. Hence, the MM estimator is asymptotically fully
efficient. In general, the asymptotic relative efficiency (ARE) of one estimator with respect to another
is defined as the ratio of their asymptotic variances. In the multivariate case, the two variances are
5
replaced by the corresponding generalized variances, which are the determinants of the asymptotic
variance-covariance matrices of the k-variate parameter estimators, and then the ratio is raised to the
power 1/k. (For further details on these issues, we refer, for example, to Serfling, 1980, Section 4.1).
Thus, for the MLE and MM estimators of the GPD parameters, we have that
ARE(MM, MLE) =(|Σ0|
/|Σ1|
)1/2=
(1 − γ)(1 + 3γ)(1 + γ)2
√(1 − 2γ)(1 + 4γ)
1 + 2γ, (2.8)
where |Σi| denotes the determinant of matrix Σi, i = 0, 1. In Table 1, we provide numerical illus-
trations of expression (2.8) for selected values of the shape parameter γ. Note that while the MLE
dominates MM for γ < 1/2, the MM continues to maintain the same (simple) asymptotic behavior for
γ ≥ 1/2, and hence it is a more practical method than the MLE in this range of γ.
Table 1: ARE(MM, MLE) for selected values of γ < 0.50.
γ (−∞;−0.25] −0.20 −0.10 −0.05 0 0.05 0.10 0.20 0.30 0.40 0.49ARE 0 .512 .902 .978 1 .982 .934 .781 .584 .362 .098
Finally, notice that the definition of the MLE and MM estimators as well as the asymptotic results
(2.6)–(2.8) are not valid over the entire parameter space. This prompts us to search for parameter-
estimation methods that work everywhere, i.e., for σ > 0 and −∞ < γ < ∞.
2.2.2 Percentile-Matching
The quantile, or percentile-matching, estimators of the GPD parameters have been proposed by
Castillo and Hadi (1997). In this section, we shall provide the computational formulas of these esti-
mators and specify their large-sample distribution.
A percentile-matching estimator of (σ, γ) of the GPD is found by matching two theoretical quantiles
F−1(p1) and F−1(p2) with the corresponding empirical quantiles Xn:[np1] and Xn:[np2], and then solving
the resulting system of equations with respect to σ and γ. Here the percentile levels p1 and p2
(0 < p1 < p2 < 1) are chosen by the researcher, the quantile function F−1(·) is given by (2.3), and [·]
denotes ‘greatest integer part’. This leads to
σ̂PM =
{γ̂PM(Xn:[np1] − x0)
/(1 − (1 − p1)bγPM), if γ̂PM 6= 0,
−(Xn:[np1] − x0)/
log(1 − p1), if γ̂PM = 0,
6
where γ̂PM is found by numerically solving the following equation with respect to γ
(Xn:[np2] − x0)/(Xn:[np1] − x0) − (F−1(p2) − x0)
/(F−1(p1) − x0) = 0.
Note that the ratio (F−1(p2) − x0)/(F−1(p1) − x0) depends only on γ and is a continuous function.
As discussed by Castillo and Hadi (1997, Section 3), the percentile-matching estimators are con-
sistent and asymptotically normal. More specifically,
(σ̂PM, γ̂PM
)∼ AN
((σ, γ), n−1Σ2
)with Σ2 = CΣ∗C
′, (2.9)
where
Σ∗ = σ2
[p1(1 − p1)2γ−1 p1(1 − p1)γ−1(1 − p2)γ
p1(1 − p1)γ−1(1 − p2)γ p2(1 − p2)2γ−1
]
and
C =1
c∗(p1, p2, γ)
[−F−1(p2) − σ(1 − p2)γ log(1 − p2) F−1(p1) + σ(1 − p1)γ log(1 − p1)
(1 − p2)γ − 1 1 − (1 − p1)γ
]
with c∗(p1, p2, γ) = F−1(p2)(1 − p1)γ log(1 − p1) − F−1(p1)(1 − p2)γ log(1 − p2).
Remark 1: The γ = 0 case.
The asymptotic normality result (2.9) is valid for −∞ < γ < ∞. When γ = 0, however, the entries
of C have to be calculated by taking the limit γ → 0. (The elements of Σ∗ are computed directly by
substituting γ = 0.) This leads to the following simplified formula of C:
C =1
log(1 − p1) log(1 − p2) log((1 − p2)/(1 − p1)
)
− log2(1 − p2) log2(1 − p1)
2 log(1 − p2)/σ −2 log(1 − p1)/σ
.
�
2.2.3 MTM Estimation
The underlying idea of the MTM—method of trimmed moments—is identical to that of the percentile-
matching, except now we match theoretical and empirical trimmed moments instead of percentiles
(quantiles). A general methodology for finding MTM estimators has been introduced and developed
by Brazauskas et al . (2009). Here we adapt their proposal to the GPD case.
7
We first calculate two sample trimmed moments (in this case, trimmed means):
µ̂j =1
n − mn(j) − m∗n(j)
n−m∗n(j)∑
i=mn(j)+1
(Xi:n − x0
), for j = 1, 2, (2.10)
where mn(j) and m∗
n(j) are integers such that 0 ≤ mn(j) < n − m∗n(j) ≤ n, and mn(j)/n → aj and
m∗n(j)/n → bj as n → ∞. The trimming proportions aj and bj must be chosen by the researcher. If
one selects aj > 0 and bj > 0 (0 < aj + bj < 1), for j = 1, 2, then the resulting estimators will be
resistant against outliers, i.e., they will be robust with the lower and upper breakdown points given
by lbp = min{a1, a2} and ubp = min{b1, b2}, respectively. The robustness of such estimators against
extremely small or large outliers comes from the fact that the order statistics with the index less than
n × lbp or more than n × (1 − ubp) are simply not included in the computation of estimates.
Next, we derive the corresponding theoretical trimmed moments:
µj =1
1 − aj − bj
∫ 1−bjaj
(F−1(u) − x0
)du
= σ ×
−1 + 11 − aj − bj
log
(1 − aj
bj
), if γ = −1,
1 +bj log(bj) − (1 − aj) log(1 − aj)
1 − aj − bj, if γ = 0,
(1/γ)
(1 −
(1 − aj)γ+1 − bγ+1j(γ + 1)(1 − aj − bj)
), otherwise.
(2.11)
Now, we match µ̂j, given by (2.10), with µj, given by (2.11), which results in a system of two equations.
After some straightforward simplifications, we find that
σ̂MTM = µ̂1 ×
−(
1 − log(1 − a1) − log(b1)1 − a1 − b1
)−1
, if γ̂MTM = −1,(
1 − (1 − a1) log(1 − a1) − b1 log(b1)1 − a1 − b1
)−1
, if γ̂MTM = 0,
γ̂MTM
(1 − (1 − a1)
bγMTM+1 − bbγMTM+11(γ̂MTM + 1)(1 − a1 − b1)
)−1
, otherwise,
where γ̂MTM is found by numerically solving the following equation with respect to γ
µ̂1/µ̂2 − µ1
/µ2 = 0.
Note that the ratio µ1/µ2 depends only on γ and is a continuous function. Also, the estimators σ̂MTM
and γ̂MTM are functions of µ̂1 and µ̂2, which we will denote g1(µ̂1, µ̂2) and g2(µ̂1, µ̂2), respectively.
8
Further, as demonstrated by Brazauskas et al . (2009), the MTM estimators are consistent and
asymptotically normal. In particular, for the GPD case, we have that
(σ̂MTM, γ̂MTM
)∼ AN
((σ, γ), n−1Σ3
)with Σ3 = DΣ∗∗D
′, (2.12)
where Σ∗∗ :=[σ2ij]2i,j=1
with the entries
σ2ij =1
(1 − ai − bi)(1 − aj − bj)
∫ 1−biai
∫ 1−bjaj
(min{u, v} − uv
)dF−1(v) dF−1(u), (2.13)
and D = [dij ]2i,j=1 with the entries dij = ∂gi/∂µ̂j
∣∣(µ1,µ2)
:
d11 = σ(d∗ − µ′1µ2)/(µ1d∗), d12 = σµ′1/d∗, d21 = µ2/d∗, d22 = −µ1/d∗.
Here d∗ = µ′
1µ2 −µ′2µ1, the trimmed moment µj is given by (2.11), and µ′j is the derivative of µj with
respect to γ, which is provided in the appendix. Of course, the entries σ2ij can be derived analytically
by performing straightforward (but tedious!) integration of (2.13), where F−1 is given by (2.3). We,
however, used the bivariate trapezoidal rule to approximate the double integral in (2.13), and found
it to be a much more effective approach than the theoretical one. It also is sufficiently accurate for all
practical purposes. The details of this approximation are presented in the appendix.
Remark 2: MTM and PM estimators.
Suppose that the percentile pj used in the PM estimation is between the aj and 1 − bj of the MTM
approach, i.e., aj < pj < 1−bj . Let aj ↑ pj and 1−bj ↓ pj. Then, it is easy to see that µj → F−1(pj)−x0and µ̂j → Xn:[npj]−x0. Hence, the MTM estimators can be reduced to the PMs. Similar computations
involving the matrices in (2.12) show that D → C, Σ∗∗ → Σ∗, and thus Σ3 → Σ2 as aj and 1 − bjapproach pj. In summary, for estimation of the GPD parameters, the PM approach is a limiting case
of the MTM. Its robustness properties also directly follow from those of the MTM. That is, the lower
and upper breakdown points of the PM estimator are: lbp = p1 and ubp = 1 − p2. �
We complete this section with an investigation of the efficiency properties of the MTM and PM
procedures with respect to the standard methods. Since the MLE is asymptotically normal (at the
n−1 rate) and fully efficient for γ < 1/2, we will use its generalized variance as a benchmark for that
range of γ. For the case γ ≥ 1/2, the MTM estimators will be compared to the MM. We will choose
9
PM and MTM estimators so that they will form pairs with respect to robustness properties. That
is, for each PM estimator with percentile levels p1 and p2, we will choose an MTM estimator with
lbp = p1 and ubp = 1 − p2. That is, the trimming proportions of the MTM estimator shall satisfy:
0 < a1 < 1 − b1 ≤ a2 < 1 − b2 < 1 and (p1, p2) = (a1, 1 − b2).
The other two proportions, a2 and 1 − b1, will be chosen to maximize the efficiency of the MTM
estimator for most of the selected γ values. Note that there are other ways to arrange the proportions
(a1, 1 − b1) and (a2, 1 − b2) while still maintaining lbp = p1 and ubp = 1 − p2. For example, we can
choose: 0 < a1 ≤ a2 < 1−b2 ≤ 1−b1 < 1 with (p1, p2) = (a1, 1−b1), or 0 < a1 ≤ a2 < 1−b1 ≤ 1−b2 < 1
with (p1, p2) = (a1, 1− b2). The advantage of the initial choice is that it directly yields PM estimators
while the other two arrangements do not. The findings of our study are summarized in Table 2.
Table 2: ARE of MTM estimators with respect to: MLE (for γ < 0.50) and MM (for γ ≥ 0.50).The PM estimators correspond to MTMs with aj ≈ pj ≈ 1 − bj, j = 1, 2, and are marked with ∗.
Trimming Proportions γ < 0.50(a1, 1 − b1) (a2, 1 − b2) −4 −2 −1 −0.40 −0.20 0 0.20 0.400.05, 0.30 0.70, 0.95 0.803 0.839 0.749 0.585 0.502 0.402 0.284 0.141
0.05∗ 0.95∗ 0.474 0.405 0.351 0.294 0.265 0.227 0.174 0.095
0.10, 0.30 0.60, 0.90 0.829 0.802 0.658 0.482 0.403 0.315 0.217 0.1050.10∗ 0.90∗ 0.648 0.562 0.472 0.373 0.326 0.268 0.196 0.102
0.15, 0.35 0.80, 0.90 0.789 0.802 0.705 0.557 0.483 0.393 0.283 0.1430.15∗ 0.90∗ 0.705 0.643 0.553 0.443 0.389 0.321 0.236 0.122
0.30, 0.50 0.70, 0.85 0.396 0.591 0.604 0.495 0.429 0.345 0.244 0.1200.30∗ 0.85∗ 0.679 0.693 0.615 0.490 0.426 0.348 0.251 0.128
0.50, 0.60 0.70, 0.75 0.295 0.391 0.374 0.299 0.258 0.208 0.148 0.0730.50∗ 0.75∗ 0.404 0.462 0.424 0.337 0.292 0.235 0.168 0.084
γ ≥ 0.500.50 0.75 1 1.50 2 2.50 3 4
0.05, 0.30 0.70, 0.95 0.412 0.495 0.614 0.981 1.607 2.674 4.501 13.1250.05∗ 0.95∗ 0.295 0.420 0.634 1.643 4.832 15.569 53.534 718.650
0.10, 0.30 0.60, 0.90 0.302 0.351 0.419 0.617 0.922 1.385 2.088 4.7830.10∗ 0.90∗ 0.305 0.397 0.540 1.088 2.384 5.572 13.704 92.599
0.15, 0.35 0.80, 0.90 0.425 0.530 0.687 1.221 2.277 4.394 8.712 36.4290.15∗ 0.90∗ 0.369 0.481 0.656 1.328 2.919 6.834 16.822 113.740
0.30, 0.50 0.70, 0.85 0.349 0.413 0.497 0.729 1.055 1.503 2.111 4.0690.30∗ 0.85∗ 0.380 0.478 0.625 1.132 2.162 4.301 8.854 40.752
0.50, 0.60 0.70, 0.75 0.215 0.259 0.321 0.508 0.813 1.301 2.079 5.2610.50∗ 0.75∗ 0.247 0.302 0.380 0.626 1.057 1.810 3.133 9.644
10
Several conclusions emerge from the table. First, note that the PM estimator with p1 = 0.50
and p2 = 0.75 is the well-known Pickands’ estimator which is highly robust but lacks efficiency. The
ARE of PM estimators can be improved by choosing p1 and p2 further apart, and hence by sacrificing
robustness. Second, for a practically relevant robustness properties, e.g., lbp ≤ 0.15 and ubp ≤ 0.10,
efficiency of the PMs can be improved by an equally robust MTM, though the improvements are not
uniform over all values of γ. Third, the least favorable range of γ for PMs and MTMs seems to be
around the exponential distribution, i.e., for γ between −0.20 and 0.40. In that range of γ, their AREs
can fall well below 0.50. Fourth, for γ ≥ 1.50, however, the PMs and MTMs perform spectacularly
with their ARE reaching even hundreds.
Remark 3: A modification of the PMs.
Castillo and Hadi (1997) also noticed that the PM estimators lack efficiency because they are based
on only two data points. Thus the information contained in other observations is not utilized. These
authors, therefore, called such estimators “initial estimates” and proposed to improve their efficiency
properties by using the following algorithm. For a sample of size n, compute PM estimates of σ and
γ for all possible pairs of percentile levels p(i)1 = i/n and p
(j)2 = j/n, 1 ≤ i < j ≤ n. Such an approach
produces(n2
)= n(n − 1)/2 estimates of σ and γ. Then, the “final estimates” are:
σ̃ = median{σ̂1, . . . , σ̂n(n−1)/2
}and γ̃ = median
{γ̂1, . . . , γ̂n(n−1)/2
}.
This modification indeed improves efficiency of the PMs, which Castillo and Hadi (1997) successfully
demonstrated using Monte Carlo simulations. The improvement, however, comes at the price of rather
inflexible robustness properties of the estimators. Also, to specify their asymptotic distribution, one
has to deal with very messy analytical derivations. We finally note that these final estimates belong
to a broad class of ‘generalized median’ estimators which, for a single-parameter Pareto model, were
extensively studied by Brazauskas and Serfling (2000). The asymptotic breakdown points of the
generalized median procedures are: ubp = 1 − lbp → 1 − 1/√
2 ≈ 0.293 as n → ∞. �
11
2.3 Simulations
Here we supplement the theoretical large-sample results of Section 2.2 with finite-sample investigations.
The objective of simulations is two-fold:
(a) to see how large the sample size n should be for the MLE, MM, PM, and MTM estimators to
achieve (asymptotic) unbiasedness and for their finite-sample relative efficiency (RE) to reach
the corresponding ARE level, and
(b) to see how robustness or non-robustness of an estimator manifests itself in computations.
To make the calculations of MM, MLE, and AREs (with respect to the MLE) possible, we will confine
our study to the values of γ in the range (−0.25; 0.50). The RE definition is similar to that of the
ARE except that we now want to account for possible bias, which we do by replacing all entries in
the variance-covariance matrix by the corresponding mean-squared errors. Also, for the objective (b),
we will evaluate the bias and RE of estimators when the underlying GPD model is contaminated. In
particular, we will employ the following ε-contamination neighborhoods:
Fε = (1 − ε)F0 + εG, (2.14)
where F0 is the assumed GPD(x0, σ, γ) model, G is a contaminating distribution which generates
observations that violate the distributional assumptions, and the level of contamination ε represents
the probability that a sample observation comes from the distribution G instead of F0. Note that the
choice ε = 0 results in a “clean” scenario which allows us to answer the questions raised in (a).
The study design is as follows: From a specified model (2.14) we generate 10,000 samples of size
n using Monte Carlo. For each sample, we estimate the scale parameter σ and the shape parameter
γ using the estimators of Section 2.2. Then we compute the average mean and RE of those 10,000
estimates. This process is repeated 10 times and the 10 average means and the 10 REs are again
averaged and their standard deviations are reported. (Such repetitions are useful for assessing standard
errors of the estimated means and REs. Hence, our findings are essentially based on 100,000 samples.)
The standardized mean that we report is defined as the average of 100,000 estimates divided by the
true value of the parameter that we are estimating. The standard error is standardized in a similar
fashion. The study was performed for the following choices of simulation parameters:
12
• Parameters of F0: x0 = 0, σ = 1, and γ = −0.20,−0.10, 0.15, 0.40.
• Distribution G: GPD with x0 = 0, σ = 1, and γ = −5.
• Level of contamination: ε = 0, 0.01, 0.05, 0.10, 0.15.
• Sample size: n = 25, 50, 100, 500.
• Estimators of (σ, γ):
– MLE and MM.
– PM with (p1, p2): (0.05, 0.95), denoted pm1; (0.15, 0.90), denoted pm2;
(0.30, 0.85), denoted pm3.
– MTM with: (a1, b1) = (0.05, 0.70) and (a2, b2) = (0.70, 0.05), denoted mtm1;
(a1, b1) = (0.15, 0.65) and (a2, b2) = (0.80, 0.10), denoted mtm2;
(a1, b1) = (0.30, 0.50) and (a2, b2) = (0.70, 0.15), denoted mtm3.
Findings of the simulation study are summarized in Tables 3, 4 and 5. Note that the entries of the
column n → ∞ in Tables 3–4 are included as target quantities and follow from the theoretical results
of Section 2.2, not from simulations.
Let us start with Tables 3 and 4 which relate to the objective (a). Several expected, as well as some
surprising, conclusions emerge from these tables. First of all, we observe that it takes quite a large
sample (n = 500) to get the bias of σ and γ within a reasonable range (e.g., within 5%) of the target.
Among the estimators under consideration and for all choices of γ, the MTMs are best behaved. But
for samples of size n ≤ 50, the bias is indeed substantial for all estimators, especially for the shape
parameter γ. Further, the MTMs’ advantage with respect to the bias criterion is even more evident
when we compare all estimators with respect to the RE (see Table 4). In many cases, MTMs nearly
attain their ARE values for n = 100. What is interesting is that REs of the MLE and MM converge to
the corresponding AREs extremely slowly when γ is near the theoretical boundaries. (Recall that the
MLE’s asymptotic normality result is valid for γ < 1/2, and the MM’s for γ > −1/4.) To get a better
idea about this issue, we performed additional simulations and discovered that, for γ = −0.20, RE of
the MM is: 0.81 (for n = 1, 000), 0.74 (for n = 2, 500), 0.70 (for n = 5, 000), 0.65 (for n = 10, 000).
13
Table 3: Standardized mean of MLE, MM, PM and MTM estimators for selected values of γ.
The entries are mean values based on 100, 000 simulated samples of size n.
γ Estimator n = 25 n = 50 n = 100 n = 500 n → ∞σ γ σ γ σ γ σ γ σ γ
−0.20 mle 1.14 0.41 1.06 0.74 1.03 0.87 1.01 0.98 1 1mm 1.16 0.28 1.10 0.56 1.06 0.73 1.02 0.91 1 1pm1 1.63 −0.49 1.20 0.74 1.21 0.43 1.04 0.89 1 1pm2 1.07 1.02 1.08 0.87 1.08 0.78 1.02 0.95 1 1pm3 1.09 0.98 1.10 0.63 1.05 0.69 1.01 0.95 1 1mtm1 1.18 0.40 1.03 1.13 1.03 0.90 1.01 0.98 1 1mtm2 1.04 1.22 1.06 0.82 1.03 0.93 1.01 0.99 1 1mtm3 1.09 0.93 1.04 1.03 1.03 0.92 1.00 0.98 1 1
−0.10 mle 1.15 −0.26 1.06 0.44 1.03 0.73 1.01 0.95 1 1mm 1.12 −0.10 1.07 0.37 1.04 0.66 1.01 0.91 1 1pm1 1.63 −2.10 1.21 0.40 1.21 −0.18 1.04 0.77 1 1pm2 1.07 1.00 1.07 0.73 1.08 0.52 1.01 0.91 1 1pm3 1.09 0.93 1.10 0.22 1.05 0.39 1.01 0.90 1 1mtm1 1.18 −0.33 1.03 1.24 1.03 0.78 1.01 0.96 1 1mtm2 1.05 1.35 1.05 0.65 1.03 0.83 1.01 0.97 1 1mtm3 1.09 0.80 1.04 1.02 1.02 0.83 1.00 0.97 1 1
0.15 mle 1.17 2.02 1.07 1.44 1.03 1.22 1.01 1.05 1 1mm 1.06 1.43 1.03 1.21 1.02 1.10 1.00 1.02 1 1pm1 1.61 3.35 1.20 1.48 1.21 1.86 1.04 1.17 1 1pm2 1.07 1.04 1.07 1.22 1.07 1.35 1.01 1.07 1 1pm3 1.08 1.14 1.09 1.56 1.05 1.40 1.01 1.07 1 1mtm1 1.17 2.00 1.03 0.91 1.03 1.16 1.01 1.03 1 1mtm2 1.04 0.84 1.05 1.26 1.02 1.12 1.01 1.03 1 1mtm3 1.08 1.22 1.04 1.04 1.02 1.12 1.00 1.02 1 1
0.40 mle 1.21 1.49 1.08 1.21 1.04 1.11 1.01 1.03 1 1mm 1.05 1.13 1.02 1.06 1.01 1.03 1.00 1.01 1 1pm1 1.61 2.06 1.20 1.25 1.21 1.38 1.04 1.07 1 1pm2 1.06 1.03 1.07 1.10 1.07 1.15 1.01 1.03 1 1pm3 1.08 1.08 1.09 1.23 1.05 1.15 1.01 1.02 1 1mtm1 1.17 1.43 1.03 0.99 1.03 1.07 1.01 1.01 1 1mtm2 1.03 0.96 1.05 1.10 1.02 1.05 1.00 1.01 1 1mtm3 1.07 1.10 1.03 1.04 1.02 1.05 1.00 1.01 1 1
Note: The ranges of standard errors for the simulated entries of σ and γ, respectively, are:
0.0001–0.0032 and 0.0007–0.0106 (for γ = −0.20); 0.0001–0.0042 and 0.0012–0.0217 (for γ = −0.10);
0.0002–0.0034 and 0.0008–0.0106 (for γ = 0.15); 0.0001–0.0052 and 0.0003–0.0092 (for γ = 0.40).
Likewise, for γ = 0.40, RE of the MLE is: 0.69 (for n = 1, 000), 0.75 (for n = 2, 500), 0.79
(for n = 5, 000), 0.82 (for n = 10, 000). Note that similar observations, with no specific numbers
though, were also made by Hosking and Wallis (1987). Finally, what is a bit surprising is that the
14
MM estimator performs quite well around γ = −0.20 and approaches its ARE from above. In typical
cases, finite-sample REs approach corresponding AREs from below (see, e.g., MLE for γ = 0.40).
Table 4: Relative efficiency of MLE, MM, PM and MTM estimators for selected values of γ.
The entries are mean values based on 100, 000 simulated samples of size n.
γ Estimator n = 25 n = 50 n = 100 n = 500 n → ∞−0.20 mle 0.61 0.79 0.89 0.98 1
mm 0.84 0.93 0.96 0.87 0.512pm1 0.13 0.18 0.19 0.24 0.265pm2 0.32 0.34 0.35 0.38 0.389pm3 0.36 0.36 0.39 0.42 0.426mtm1 0.37 0.46 0.47 0.50 0.502mtm2 0.43 0.44 0.46 0.48 0.483mtm3 0.36 0.39 0.40 0.41 0.429
−0.10 mle 0.56 0.76 0.87 0.97 1mm 0.86 0.96 1.00 0.97 0.902pm1 0.11 0.16 0.18 0.23 0.247pm2 0.29 0.31 0.32 0.35 0.357pm3 0.33 0.33 0.36 0.38 0.389mtm1 0.33 0.42 0.43 0.45 0.454mtm2 0.39 0.40 0.42 0.44 0.440mtm3 0.33 0.35 0.36 0.38 0.389
0.15 mle 0.41 0.63 0.75 0.91 1mm 0.66 0.76 0.81 0.85 0.865pm1 0.08 0.11 0.12 0.17 0.189pm2 0.20 0.22 0.23 0.25 0.259pm3 0.23 0.24 0.25 0.27 0.277mtm1 0.23 0.29 0.29 0.31 0.315mtm2 0.28 0.28 0.30 0.31 0.312mtm3 0.23 0.24 0.25 0.26 0.271
0.40 mle 0.18 0.33 0.44 0.62 1mm 0.28 0.32 0.35 0.36 0.362pm1 0.03 0.05 0.06 0.08 0.095pm2 0.09 0.10 0.10 0.12 0.122pm3 0.11 0.11 0.11 0.13 0.128mtm1 0.10 0.13 0.13 0.14 0.141mtm2 0.13 0.13 0.14 0.14 0.143mtm3 0.11 0.11 0.11 0.12 0.120
Note: The range of standard errors for the simulated entries is:
0.0007–0.0036 (for γ = −0.20); 0.0007–0.0042 (for γ = −0.10);
0.0005–0.0028 (for γ = 0.15); 0.0002–0.0025 (for γ = 0.40).
In Table 5, we illustrate the behavior of estimators under several data-contamination scenarios.
We choose GPD(x0 = 0, σ = 1, γ = 0.15) as the “clean” model because for γ = 0.15 the MLE
15
and MM estimators are much more efficient than the PM and MTM estimators (see Table 4). As
one can see from Table 5, however, just 1% of “bad” observations can completely erase the huge
advantage of standard (non-robust) procedures over the robust PMs and MTMs. Indeed, for ε > 0, the
MLE estimates become totally uninformative, and the MM procedure simply collapses. On the other
hand, the robust estimators stay on target when estimating σ and exhibit a gradually deteriorating
performance as ε increases. The deterioration, in this case, is not unexpected for as the level of data-
contamination reaches or exceeds PMs’ and MTMs’ ubp, they also become uninformative. Finally,
note that for all estimators the primary source of impaired performance is the bias in estimation of γ.
Table 5: Mean and relative efficiency of MLE, MM, PM and MTM estimators under several
data-contamination models Fε = (1 − ε)GPD(x0 = 0, σ = 1, γ = 0.15) + εGPD(σ = 1, γ = −5).The entries are mean values based on 100, 000 simulated samples of size n = 500.
Statistic Estimator ε = 0 ε = 0.01 ε = 0.05 ε = 0.10 ε = 0.15σ̂ γ̂ σ̂ γ̂ σ̂ γ̂ σ̂ γ̂ σ̂ γ̂
mean mle 1.01 0.16 0.76 −0.29 0.65 −0.81 0.61 −1.19 0.60 −1.51mm∗ 1.00 0.15 ∞ −0.41 ∞ −0.50 ∞ −0.50 ∞ −0.50pm1 1.04 0.18 1.04 0.15 1.04 0.05 1.04 −0.46 1.02 −1.36pm2 1.01 0.16 1.02 0.14 1.02 0.07 1.03 −0.05 1.03 −0.25pm3 1.01 0.16 1.01 0.15 1.02 0.08 1.03 −0.01 1.03 −0.12mtm1 1.01 0.15 1.01 0.14 1.01 0.07 1.01 −0.12 0.96 −0.74mtm2 1.01 0.15 1.01 0.14 1.01 0.07 1.02 −0.02 1.02 −0.15mtm3 1.01 0.15 1.01 0.14 1.02 0.08 1.03 0.00 1.03 −0.10
re mle 0.91 0.07 0.03 0.02 0.02mm∗ 0.85 0 0 0 0pm1 0.17 0.14 0.05 0.01 0.00pm2 0.25 0.23 0.10 0.05 0.02pm3 0.27 0.26 0.14 0.07 0.04mtm1 0.31 0.28 0.13 0.04 0.02mtm2 0.31 0.28 0.14 0.07 0.04mtm3 0.26 0.25 0.15 0.08 0.05
∗ For the MM estimator, the entries ∞ and 0 correspond to numbers of the order 1021 and 10−20, respectively.
3 Actuarial Applications
In this section, we fit the GPD model to the Danish insurance data which has been extensively studied
in the actuarial literature (see, e.g., McNeil, 1997). We also investigate the implications of a model fit
on risk evaluations and ratemaking. In particular, we use empirical, parametric and robust parametric
16
approaches to compute point estimates of several value-at-risk measures and net premiums of a few
insurance contracts.
3.1 Fitting Insurance Data
The Danish insurance data were collected at Copenhagen Re and comprise 2167 fire losses over the
period 1980 to 1990. They have been adjusted for inflation to reflect 1985 values and are expressed
in millions of Danish krones (dkk). All losses are one million dkk or larger. A thorough diagnostic
analysis of this data set was performed by McNeil (1997) who concluded that the GPD assumption
is a reasonable one for the Danish insurance data. He then used the MLE approach to fit the GPD
to the whole data set as well as to data above various thresholds. The important issues of ‘model
uncertainty’, ‘parameter uncertainty’, and ‘data uncertainty’ were also discussed by McNeil (1997).
Some of those discussions are taken as motivation for the foregoing investigations.
In this section, we fit the GPD to the entire data set and to data above several thresholds, and
examine the stability of fits under a few data-perturbation scenarios. Our main objectives are to see:
(a) how important is the choice of a parameter-estimation method on the model fit, and (b) what
influence it has on subsequent pricing and risk measurement exercises. For visual and quantitative
assessments of the quality of model fit, we employ the percentile-residual (PR) plot and a trimmed
mean absolute deviation (tMAD), respectively. These tools are taken from Brazauskas (2009) and are
defined as follows. The PR plots are constructed by plotting the empirical percentile levels, (j/n)100%,
versus the standardized residuals
Rj,n =Xj:n − F̂−1
(j−0.5
n
)
standard deviation of F̂−1(
j−0.5n
) for j = 1, . . . , n, (3.1)
where Xj:n is the observed (j/n)th quantile and the qf F−1, given by (2.3), is estimated by replacing
parameters σ and γ with their respective estimates σ̂ and γ̂. For finding σ̂ and γ̂ we use the following
estimators: MLE, PM3 with (p1, p2) = (0.30, 0.85), MTM3 with (a1, b1) = (0.30, 0.50), (a2, b2) =
(0.70, 0.15), and MTM4 with (a1, b1) = (0.10, 0.55), (a2, b2) = (0.70, 0.05). The MM approach is
excluded from further consideration because its asymptotic properties are not valid for the Danish
insurance data. The denominator of (3.1) will be estimated by using the delta method (see, e.g.,
Serfling, 1980, Section 3.3) in conjunction with the corresponding variance-covariance matrix Σ0, Σ2,
17
or Σ3; the matrices are defined by (2.6), (2.9), (2.12), respectively. In the PR-plot, the horizontal line
at 0 represents the estimated quantiles, and the ±2.5 lines are the tolerance limits. A good fit would
be the one for which the majority of points (ideally, all points) are scattered between the tolerance
limits. The PR-plots for MLE, PM, and MTM fits are presented in Figure 2. The plots are based on
the whole data set, i.e., on 2156 losses in excess of one million dkk.
0 10 20 30 40 50 60 70 80 90 100
−6
−4
−2
0
2
4
6
Empirical Percentile Levels
Sta
nd
ard
ize
d
Re
sid
ua
ls
MLE Fit
0 10 20 30 40 50 60 70 80 90 100
−6
−4
−2
0
2
4
6
Empirical Percentile Levels
Sta
nd
ard
ize
d
Re
sid
ua
ls
PM3 Fit
0 10 20 30 40 50 60 70 80 90 100
−6
−4
−2
0
2
4
6
Empirical Percentile Levels
Sta
nd
ard
ize
d
Re
sid
ua
ls
MTM3 Fit
0 10 20 30 40 50 60 70 80 90 100
−6
−4
−2
0
2
4
6
Empirical Percentile Levels
Sta
nd
ard
ize
d
Re
sid
ua
ls
MTM4 Fit
Figure 2: PR-plots for the GPD model fitted by the MLE, PM, and MTM methods.
As one can see from Figure 2, all parameter-estimation methods do a mediocre job at fitting the
“small” losses, i.e., those slightly above the threshold of one million, but they perform reasonably well
18
for the “medium” and “large” section of the losses. Overall, among the four PR-plots, the MLE fit
looks worst. One should keep in mind, however, that the vertical deviations in these plots depend on
the efficiency of the estimator. Thus, the same value of absolute residual will appear as significantly
larger on the MLE’s plot than it will on, for example, the MTM4’s because the latter estimator is
less efficient. Indeed, its ARE is: 0.832 (for γ = −1), 0.711 (for γ = −0.50), 0.609 (for γ = −0.25).
Therefore, since for practical decision-making the actual (not statistical!) discrepancies matter more,
it is important to monitor the non-standardized residuals as well.
Next, the tMAD measure evaluates the absolute distance between the fitted GPD quantiles and
the observed data. It is defined by ∆δ =1
[nδ]
∑[nδ]i=1 bi:n, where bi:n denotes the ith smallest distance
among∣∣Xj:n − F̂−1 ((j − 0.5)/n)
∣∣, j = 1, . . . , n. We use the following values of δ: 0.50, 0.75, 0.90,
0.95, 1. The choice δ = 0.90, for instance, indicates how far, on the average, are the 90% closest
observations from their corresponding fitted quantiles. In Table 6, we report parameter estimates and
the goodness-of-fit measurements ∆δ for various data thresholds x0.
Table 6: Parameter estimates and goodness-of-fit measurements of the GPD model
for selected model-fitting procedures and several data thresholds x0.
Threshold Fitting Parameter Estimates Model Fit (∆δ)(Excesses) Procedure σ̂ γ̂ δ = 0.50 δ = 0.75 δ = 0.90 δ = 0.95 δ = 1
x0 = 1 mle 0.946 −0.604 0.02 0.04 0.05 0.06 0.19(2156) pm3 1.036 −0.501 0.01 0.03 0.05 0.07 0.43
mtm3 0.989 −0.520 0.02 0.03 0.04 0.07 0.41mtm4 1.035 −0.515 0.01 0.04 0.05 0.07 0.39
x0 = 3 mle 2.189 −0.668 0.03 0.06 0.16 0.23 0.74(532) pm3 2.171 −0.788 0.03 0.10 0.17 0.29 2.34
mtm3 2.079 −0.794 0.03 0.08 0.12 0.22 2.19mtm4 2.209 −0.720 0.03 0.10 0.16 0.21 1.26
x0 = 10 mle 6.975 −0.497 0.12 0.26 0.46 0.61 2.21(109) pm3 7.101 −0.345 0.13 0.28 0.41 0.64 3.53
mtm3 7.819 −0.290 0.08 0.15 0.24 0.47 3.51mtm4 7.546 −0.377 0.08 0.18 0.33 0.43 2.85
x0 = 20 mle 9.635 −0.684 0.28 0.52 0.91 1.34 3.32(36) pm3 11.751 −0.476 0.29 0.70 1.12 2.51 6.27
mtm3 9.920 −0.686 0.31 0.64 1.13 1.80 3.59mtm4 10.524 −0.813 0.37 1.33 2.73 4.06 9.13
After examining Table 6 we make the following observations. Clearly, the MLE fits are best for
19
all data-thresholds under consideration if we measure the fit by ∆δ with δ = 1. This should not be
surprising since likelihood-based procedures attempt, and are designed, to fit all data points. But if
we look at the other values of δ (which reflect the fit for most—not all—observations), we see that
the MLE and robust fits are similar and fairly close to the actual data, for all data-thresholds. We
also note a strong performance by MTMs for x0 = 10. Aside from the last point, however, so far
the robust procedures have not offered any significant improvements over the MLE. But this changes
substantially when we perform a sensitivity analysis under a few data-perturbation scenarios.
Table 7: Parameter estimates and goodness-of-fit measurements of the GPD model
for selected model-fitting procedures, x0 = 10, and under several data-perturbation scenarios.
Scenario Fitting Parameter Estimates Model Fit (∆δ)Procedure σ̂ γ̂ δ = 0.50 δ = 0.75 δ = 0.90 δ = 0.95 δ = 1
Remove mle 7.230 −0.390 0.09 0.19 0.30 0.40 1.14x = 263 pm3 7.132 −0.321 0.11 0.23 0.35 0.47 1.66
mtm3 7.709 −0.267 0.08 0.15 0.24 0.38 1.73mtm4 7.420 −0.336 0.08 0.16 0.26 0.35 1.32
Add mle 6.778 −0.598 0.16 0.32 0.68 0.90 3.70x = 350 pm3 7.422 −0.304 0.11 0.29 0.49 0.90 6.70
mtm3 7.897 −0.316 0.09 0.16 0.26 0.56 6.09mtm4 7.620 −0.421 0.09 0.21 0.40 0.53 5.14
Replace mle 6.892 −0.517 0.13 0.26 0.51 0.68 3.02x = 263 pm3 7.101 −0.345 0.13 0.28 0.41 0.64 4.42
with mtm3 7.819 −0.290 0.08 0.15 0.24 0.47 4.40x = 350 mtm4 7.546 −0.377 0.08 0.18 0.33 0.43 3.74
In Table 7, we report parameter estimates and the goodness-of-fit measurements ∆δ for x0 = 10
under the following three scenarios. The first two scenarios are taken from McNeil (1997) and the
third one is a combination of the other two. In the first scenario (labeled “Remove x = 263”) we
remove the largest loss from the original sample. In the second scenario (labeled “Add x = 350”) we
introduce a new largest observation of 350 to the data set. And in the third scenario (labeled “Replace
x = 263 with x = 350”) we replace the current largest point of 263 with a new loss of 350. If we
compare parameter estimates and ∆δ evaluations with the original results in Table 6 (for x0 = 10),
we see that robust estimates and their fits are significantly less affected by the removal or addition of
the largest loss than those of the MLE. Moreover, for the third scenario all robust estimates and their
fit measurements for δ < 1 are absolutely identical to the original ones whilst the results for MLE
20
are distorted. In the next two sections, we further explore the data-perturbation effects in insurance
applications by studying their influence on the estimates of risk measures and net premiums.
3.2 Risk Measurement
To see how the quality of model fit affects insurance risk evaluations, we compute empirical, parametric
and robust parametric point estimates for several value-at-risk (VaR) measures. The computations are
performed for the whole data set (i.e., for x0 = 1) and for data above the threshold x0 = 10 including
some data-perturbation scenarios of Section 3.1. Mathematically, the VaR measure is the (1−β)-level
quantile of the distribution function F , that is, VaR(β) = F−1(1 − β). For empirical estimation, we
replace F with the empirical cdf F̂n and arrive at
V̂aREMP(β) = Xn:n−[nβ].
For parametric (MLE) and robust parametric (PM, MTM) estimation, F̂−1 is found by simply replac-
ing parameters σ and γ with their respective MLE, PM, and MTM estimates in (2.3). Note that for
parametric quantile-estimation based on upper tail of the data (i.e., for the thresholds x0 > 1), we
apply the results of McNeil (1997, Section 3.5). In particular, we use σ, γ and location x0 estimates
which are calculated according to the following formulas:
γ̃ = γ̂, σ̃ = σ̂(1 − F̂n(x0))−eγ , x̃0 = x0 + (σ̃/γ̃)((1 − F̂n(x0))eγ − 1
),
where σ̂ and γ̂ are the parameter estimates based only on data above x0. Table 8 presents empirical,
parametric, and robust parametric point estimates of VaR(β) for several levels of β.
A number of conclusions emerge from the table. First, for the whole data set (x0 = 1), GPD-based
risk evaluations of not-too-extreme significance levels (β ≥ 0.01) are fairly close to their empirical
counterparts. Second, for very extreme significance levels (β < 0.01), the empirical and parametric
estimates of VaR diverge. Of course, one can argue that the empirical approach underestimates
VaR(β = 0.0001) because there is simply no observed data at that level. On the other hand, the
MLE’s estimate seems like an exaggeration of risk. Third, the last point gains even more credibility
if we look at VaR(β < 0.01) estimates which are based on data above x0 = 10. Indeed, the MLE’s
evaluations are now substantially reduced. Fourth, overall the robust procedures tend to provide lower
21
estimates of risk at the most extreme levels of significance than the MLE. This actually can easily
be seen from the PR-plots (see Figure 2): near the 100th percentile, MLE’s residuals are below the
fitted line whilst the PM’s and MTM’s are above. Fifth, when we employ data-perturbation scenarios,
robust estimators’ risk evaluations are quite stable compared to those of the MLE.
Table 8: Point estimates of various value-at-risk measures computed by employing
empirical, parametric (MLE), and robust parametric (PM and MTM) methods.
Scenario Estimation VaR(β)Method β = 0.10 β = 0.05 β = 0.01 β = 0.001 β = 0.0001
All Data mle 5.73 9.0 25 101 408(x0 = 1) pm3 5.48 8.2 20 65 207
mtm3 5.40 8.1 20 68 228mtm4 5.57 8.4 21 70 230
empirical 5.56 10.1 26 145 263
x0 = 10 mle 5.96 10.1 27 95 306pm3 5.68 10.1 25 69 166
mtm3 5.16 10.1 26 67 147mtm4 5.46 10.1 27 78 199
x0 = 10 mle 6.04 10.1 27 98 331and replace pm3 5.68 10.1 25 69 166
x = 263 mtm3 5.16 10.1 26 67 147with x = 350 mtm4 5.46 10.1 27 78 199
x0 = 10 mle 6.20 10.1 29 117 469and add pm3 5.43 10.1 26 66 147x = 350 mtm3 5.15 10.1 27 71 164
mtm4 5.48 10.1 28 86 241
3.3 Contract Pricing
Consider now estimation of the pure premium for an insurance benefit equal to the amount by which
a loss exceeds l (million dkk) with a maximum benefit of m. That is, if the fire damage is X with
distribution function F , we seek
Π[F ] =
∫ l+m
l(x − l) dF (x) + m
(1 − F (l + m)
). (3.2)
Since Π[F ] is a functional of the underlying loss distribution F , we can estimate it by replacing F with
its estimate. To accomplish that, we employ three approaches: empirical, parametric (MLE), and
22
robust parametric (PM and MTM). In addition, we also provide (estimated) standard errors of the
premium Π[F ] estimates. To find parametric estimates of the errors, we use the delta method applied
to the transformation of parameter estimators given by equation (3.2) together with the MLE, PM,
and MTM asymptotic distributions, which have been discussed earlier. For the empirical estimation
of standard errors, we use the classical central limit theorem and have that
Π[F̂n] ∼ AN(Π[F ], n−1V [F ]
),
where V [F ] =∫ l+ml (x−l)2 dF (x)+m2(1−F (l+m))−(Π[F ])2 and F̂n denotes the empirical distribution
function. Further, the reliability of premium estimates is studied by employing two data-perturbation
scenarios. In the first scenario (labeled “Replace Top 10 with ∼ 350”), we make the ten largest losses
even larger by replacing them with 351, 352, . . . , 360. In the second scenario (labeled “Replace Top
10 with ∼ 100”), we replace the ten largest losses with 101, 102, . . . , 110. Table 9 summarizes our
numerical investigations for contract pricing.
Table 9: Empirical, parametric (MLE), and robust parametric (PM and MTM) point
estimates of Π[F ], for selected insurance contracts and under two data-perturbation scenarios.
Estimated standard errors of the premium estimates are presented in parentheses.
Scenario Estimation Insurance contract specified by (l,m)
Method (l,m) = (2, 3) (l,m) = (5, 10) (l,m) = (20, 20) (l,m) = (50, 50)Premium Premium Premium Premium
Original mle 0.69 (0.021) 0.51 (0.037) 0.16 (0.025) 0.09 (0.021)Data pm3 0.69 (0.024) 0.44 (0.051) 0.10 (0.032) 0.04 (0.021)
(x0 = 1) mtm3 0.67 (0.023) 0.43 (0.056) 0.10 (0.035) 0.05 (0.024)mtm4 0.70 (0.023) 0.46 (0.043) 0.11 (0.027) 0.05 (0.019)
empirical 0.66 (0.023) 0.54 (0.043) 0.17 (0.034) 0.08 (0.041)
Replace mle 0.70 (0.021) 0.55 (0.038) 0.19 (0.028) 0.12 (0.026)Top 10 pm3 0.69 (0.024) 0.44 (0.051) 0.10 (0.032) 0.04 (0.021)with mtm3 0.67 (0.023) 0.43 (0.056) 0.10 (0.035) 0.05 (0.024)∼ 350 mtm4 0.70 (0.023) 0.46 (0.043) 0.11 (0.027) 0.05 (0.019)
empirical 0.66 (0.023) 0.54 (0.043) 0.17 (0.034) 0.08 (0.041)
Replace mle 0.69 (0.021) 0.52 (0.037) 0.17 (0.026) 0.09 (0.022)Top 10 pm3 0.69 (0.024) 0.44 (0.051) 0.10 (0.032) 0.04 (0.021)with mtm3 0.67 (0.023) 0.43 (0.056) 0.10 (0.035) 0.05 (0.024)∼ 100 mtm4 0.70 (0.023) 0.46 (0.043) 0.11 (0.027) 0.05 (0.019)
empirical 0.66 (0.023) 0.54 (0.043) 0.17 (0.034) 0.08 (0.041)
23
We observe that premium estimates for the contract (l,m) = (2, 3), which covers events that
are quite likely but have relatively low economic impact, are practically identical. And the data-
perturbation scenarios have no effect on any of the estimators, including the MLE. For more ex-
treme coverages, i.e., for contracts (l,m) = (5, 10) and higher, which cover low-probability-but-high-
consequence events, the estimates diverge. The MLE and empirical estimates are usually close to
each other whilst the robust estimates form their own cluster which yields fairly different prices
from those of the MLE and empirical methods. Further, for very extreme layers of losses such as
(l,m) = (20, 20), (50, 50), the parametric estimators produce smaller standard errors than the empir-
ical approach. Finally, since both data contaminations occur outside the defined layers, one would
assume that the premium should not change. This property is exhibited by the empirical and robust
approaches, but the MLE is clearly affected by contamination.
4 Discussion
In this article, we have introduced a new method for robust fitting of the GPD. It is based on ‘trimmed
moments’ and therefore called the method of trimmed moments (MTM). Its large- and small-sample
properties have been explored and compared to some well-established standard approaches (MLE
and MM) as well as to some less-known but conceptually sound methods such as ‘percentile match-
ing’ (PM). We have found that the new MTM procedure is computationally attractive, it possesses
competitive efficiency properties and provides sufficient protection against various data contamination
sources. A connection between the MTMs and PMs has also been established. Thus it is not surprising
at all that these two procedures have some desirable properties in common. In particular, they offer
a variety of robustness-efficiency trade-offs and, unlike other existing proposals in the literature, both
are applicable and valid for the entire parameter space of the GPD (i.e., for −∞ < γ < ∞).
In addition, the favorable theoretical and computational properties of the new method translate
into accurate risk evaluations as well as fair pricing. Indeed, as it is confirmed by our numerical
illustrations, the value-at-risk estimates based on the robust procedures show more stability under
various scenarios of data-perturbation than those based on the MLE. Also, robustly estimated contract
prices for extreme layers of losses are less volatile than the empirical ones, and more outlier resistant
24
than the MLE-based prices.
5 Acknowledgment
The first author gratefully acknowledges the support provided by a grant from the Actuarial Profession
(United Kingdom).
A Appendix: Auxiliary Results
A.1 Differentiation of µi
The derivative of a trimmed moment µj , given by (2.11), with respect to the shape parameter γ is
necessary for computations of the variance-covariance matrix Σ3, given by (2.12). The general case is
found by straightforward differentiation. The special cases γ = −1 and γ = 0 represent the limit of
the general case when γ → −1 and γ → 0, respectively. Thus, we have:
µ′j = −σ ×
1 − log(1 − aj) − log(bj)1 − aj − bj
− log2(1 − aj) − log2(bj)2(1 − aj − bj)
, if γ = −1,
1 − (1 − aj) log(1 − aj) − bj log(bj)1 − aj − bj
+(1 − aj) log2(1 − aj) − bj log2(bj)
2(1 − aj − bj), if γ = 0,
1
γ(γ + 1)
[(2γ + 1)(µj/σ) − 1 +
(1 − aj)γ+1 log(1 − aj) − bγ+1j log(bj)1 − aj − bj
], otherwise.
A.2 Approximation of σ2ij
The entries σ2ij appear in the variance-covariance matrix Σ∗∗ which is a part of Σ3, given by (2.12).
Instead of pursuing exact formulas for these entries, we use the bivariate trapezoidal rule.
Let us define the region R = [a, b] × [c, d] which is a subset of [0, 1) × [0, 1). Next, we divide each
interval, [a, b] and [c, d], into k subintervals and define
hx = (b − a)/k, xi = a + ihx, i = 0, . . . , k,
hy = (d − c)/k, yj = c + jhy , j = 0, . . . , k,
where the number k is fixed.
The double integral over R to be approximated is of the form
I =
∫ b
a
∫ d
cg(x, y)f ′(x)f ′(y) dy dx.
25
Using the composite trapezoidal rule in both spatial directions yields
I ≈ Ik =k∑
i=0
k∑
j=0
g(xi, yj) f′(xi)f
′(yj)wiwj ,
where
wi = hx ×{
1/2, if i = 0 or i = k,1, otherwise,
and wj = hy ×{
1/2, if j = 0 or j = k,1, otherwise.
If the integrand g(·, ·) f ′(·)f ′(·) is sufficiently smooth, then the order of convergence is O(h2).
The derivatives f ′(x) and f ′(y) are approximated by
f ′(xi) ≈1
2hx
(f(xi+1) − f(xi−1)
), i = 1, . . . , k,
f ′(yj) ≈1
2hy
(f(yj+1) − f(yj−1)
), j = 1, . . . , k,
where xi, yj, hx and hy are given above. Here, if the function f is assumed to be sufficiently smooth,
then the order of convergence is O(h2). This means that the error-of-convergence ratios,
eoc =I − Ik/2I − Ik
,
approach 4 if the true solution is known. If not, then the following ratios
eoc ≈Ik − Ik/2
Ik/2 − Ik/4
approach 4.
Finally, note that all the smoothness conditions made above are satisfied by the GPD.
References
[1] Brazauskas, V. (2009). Robust and efficient fitting of loss models: diagnostic tools and insights.
North American Actuarial Journal (in press).
[2] Brazauskas, V. and Serfling, R. (2000). Robust and efficient estimation of the tail index of
a single-parameter Pareto distribution (with discussion). North American Actuarial Journal ,
4(4), 12–27. Discussion: 5(3), 123–126. Reply: 5(3), 126–128.
[3] Brazauskas, V., Jones, B.L., and Zitikis, R. (2009). Robust fitting of claim severity distributions
and the method of trimmed moments. Journal of Statistical Planning and Inference, 139(6),
2028–2043.
26
[4] Castillo, E. and Hadi, A.S. (1997). Fitting the generalized Pareto distribution to data. Journal
of the American Statistical Association, 92, 1609–1620.
[5] Cebrian, A.C., Denuit, M., and Lambert, P. (2003). Generalized Pareto fit to the Society of
Actuaries’ large claims database. North American Actuarial Journal , 7(3), 18–36.
[6] Davison, A.C. and Smith, R.L. (1990). Models for exceedances over high thresholds (with
discussion). Journal of the Royal Statistical Society (Series B), 52, 393–442.
[7] Dell’Aquila, R. and Embrechts, P. (2006). Extremes and robustness: a contradiction? Financial
Markets and Portfolio Management , 20(1), 103–118.
[8] Dupuis, D.J. (1999). Exceedances over high thresholds: a guide to threshold selection. Extremes,
1(3), 251–261.
[9] He, X. and Fung, W.K. (1999). Method of medians for lifetime data with Weibull models.
Statistics in Medicine, 18, 1993–2009.
[10] Hosking, J.R.M. and Wallis, J.R. (1987). Parameter and quantile estimation for the generalized
Pareto distribution. Technometrics, 29(3), 339–349.
[11] Johnson, N.L., Kotz, S., and Balakrishnan, N. (1994). Continuous Univariate Distributions,
2nd edition; Volume 1. Wiley, New York.
[12] Juárez, S.F. and Schucany, W.R. (2004). Robust and efficient estimation for the generalized
Pareto distribution. Extremes, 7(3), 237–251.
[13] Klugman, S.A., Panjer, H.H., and Willmot, G.E. (2004). Loss Models: From Data to Decisions,
2nd edition. Wiley, New York.
[14] McNeil, A.J. (1997). Estimating the tails of loss severity distributions using extreme value
theory. ASTIN Bulletin, 27, 117–137.
[15] McNeil A.J., Frey R., and Embrechts P. (2005). Quantitative Risk Management: Concepts,
Techniques and Tools. Princeton University Press.
[16] Peng, L. and Welsh, A.H. (2001). Robust estimation of the generalized Pareto distribution.
Extremes, 4(1), 53–65.
[17] Serfling, R.J. (1980). Approximation Theorems of Mathematical Statistics. Wiley, New York.
[18] Smith, R.L. (1985). Maximum likelihood estimation in a class of nonregular cases. Biometrika,
72, 67–90.
27